Monitoring Sentiment: Extending Latent Dirichlet Allocation to a Hierarchical Bayesian State Space Model

In an age of modern technology, attempting to understand change in sentimentin target populations through the lens of social media is becoming both complex and commonplace. This study seeks to visually quantify change in sentiment of English speaking X users (formally know as Twitter) with regards to the current Russia-Ukraine conflict. A Bayesian hierarchical model is presented in which one of the hyper-parameters is made to be time varying using a state space model specification. Gibbs Sampling is used as the approximation method during the posterior inference step of the Latent Dirichlet Allocation algorithm and Forward Filtering Backward Sampling is used inside of Gibbs sampling when drawing for the state variable, the hyper-parameter that was made time varying. Examples of the full state space model with X data along with a special case scenario (where the time varying hyper-parameter is not Markov, but simply iid) using both X data and simulated data are given with discussion.